Conan zero-world origin investigation, 2026-11-06

No emulator mismatch was demonstrated. A fresh generated native full-shell control captured the selected first stores correctly, but rejected the existing lifetime observer on overhead: 2,479.56 ms versus 241.79 ms median for the same fixed guest workload, a 10.2551 ratio. The predeclared maximum was 1.20. No original Conan session was launched, no emulator repair was made, and movement or a second gameplay response remains unverified.

Evidence is in scratch/runs/20261106-conan-zero-world-origin/, especially control-summary.json, native-work-timing.json, observer-qualification.json, the two command-controls-*.json receipts, and the generated control captures. Source recovery uses Git 4ea0bc8afb37932dfa932c91ee06c861e0f1d06c, relevant.binarydiff, relevant untracked-file patches, and run-helper patches; there is no source-tree or media archive. The later offline summary has its own patch and hash. The explicit coordinator grant was 20:12:55–20:57:55 UTC, including cleanup, after ordered Aranna and legacy-combined FINAL releases and a fresh independent process/port/profile/lock check.

The shipped loading evidence remains the concrete original boundary: at 19:30:55 UTC, endpoint 1, device 136454160, generation 1, sequence 71332 first failed the strict singular-normal guard. The first twelve world values were zero while translation was intact. Writer serial 114805, thread 1, guest EIP 0x41ccee, return 0x41cd09, and source guest pointer 0x074ff2d0 associate the exact source bits with the copied, lowered and executed world matrix. sourceOrigin is null. Earlier construction stores for that particular source and lifetime remain unobserved.

Read-only review of the earlier producer investigations found correctly copied zero input at the final matrix copy, native-matching x87 products, and zero channel weights causing the observed palette scaling stores. These older owners and diagnostic modules do not establish sequence 71332's lifetime or a CPU defect. Nor does their loading duration establish observer overhead for the shipped original. No guest bytes, matrices, guard behavior, CPU/timer semantics, or held operations were changed or retried.

The new control is an ordinary generated Win32 PE launched through the full browser shell. Its native key callback runs one constructor, 200 time/weight updates, and exactly 15,000,000 add/store iterations. The visible success title requires the computed counter to equal 45,000,000. A separate native click callback changes the title again. Each input used the accepted serialized consumer, a personally viewed current capture and matching review hash, and verified live page/owning-CPU module identity. Both module sessions passed all 14 rejection cases before input and retained balanced trusted press/release events. These are generated control responses, not Conan gameplay.

The observer retained the actual first pre/post pairs for generated owner 0x409000, endpoint 1, worker slot 0, thread 1, epoch 1:

Guest PC Owner offset Before → after bits Serial pair
0x431955 +0x04 3f800000 → 00000000 3/4
0x431958 +0x38 3f800000 → 00000000 5/6
0x43195b +0x68 3f800000 → 00000000 7/8
0x4310b0 +0x38 00000000 → 3f000000 14/15

The final generated epoch retained all 201 calls and 1,407 details, with no open call, invalid history, incomplete packet, observer error or truncation. Two reconstructed prior epochs also closed cleanly. Limits stayed at 16 resident lifetimes, 2,048 calls and 8,192 details. Completed owning-worker slices recorded micro-op execution of the large loop. Selected-writer native entry markers remained zero, so this does not prove complete history for a selected writer executing through every hot native path. The previously documented native-destructor history gap remains a limitation.

Timing excludes module startup and measures trusted Enter to the native computed success title, with a 10 ms polling bound. Each module had a discarded warm-up and two measured repetitions:

Module Warm-up ms Measured ms Median ms
Shipped baseline 1,482.63 271.26 / 212.32 241.79
Lifetime observer 2,313.84 2,438.30 / 2,520.83 2,479.56

These are separate sequential sessions with two measured repetitions, not an alternating benchmark or a Conan timing-invariance claim. Matching source, generated input, work count, architectural success and input plumbing make the large observed cost sufficient to reject this observer for the next original. Completed-slice guest times and process CPU captures are retained separately; those capture intervals include the command's passive wait.

Validation passed: helper syntax, required locked full build, same-source standard and compatibility baseline/diagnostic compilation, byte equality of both baselines with shipped modules, unchanged diagnostic import/export contracts, and both native input/control suites. Only the standard modules ran the new browser workload. Standard baseline SHA-256 is 280014368b95a768bd31397971715f5d95089c3a22d018758b5d58f3d4a1650e; diagnostic is e895153177169d2ac4da4a816e26d37d6013c1b6e4ef91f3125704f22341a44d. The overhead and combined observer qualification gates failed as intended. Current-source REP, SEH, nesting, full passivity and singular-render controls were not rerun after this rejection and are not claimed qualified.

Both browsers, servers and profiles completed ordinary cleanup, and the runtime holder released its locks after an independent cleanup PASS. Final owned process/port/profile/lock cleanup and publication are recorded in the run. No boxes were used. The next useful work is a lower-cost observer that passes generated first-store and hot-path coverage controls before another explicitly authorized original session, then traces the actual sequence-keyed source's construction stores. A CPU or timer fix is not supported by this evidence.